First-Time Quick Start
When training and running a model for the first time, follow these five steps:
- Complete device setup and verification. Open SenseCraft Robotics, click Get Started on the home page, select the device model, and bind the leader arm, follower arm, and cameras in sequence. Then complete robotic arm calibration and teleoperation verification. Confirm that the follower arm follows the leader arm steadily in the same direction. If visual training is required, also confirm that the cameras clearly capture the workspace and task objects.
- Plan and collect demonstration data. Go to Collection, define one specific task, and enter the dataset name, task description, number of episodes, episode duration, and rest duration. During recording, use the leader arm to guide the follower arm through the complete task, while keeping the robotic arm configuration, camera positions, and task objective consistent. The platform generates a raw dataset when recording is complete.
- Inspect and organize the dataset. Go to Dataset and check the videos, sample frames, episode durations, frame rates, and joint curves. Delete episodes with failures, collisions, severe occlusion, or obvious frame drops. Data collected in separate batches may be merged only when the task and device configuration are consistent. Proceed to training only after confirming that data quality is acceptable.
- Train and obtain a model. Go to Training, select an inspected dataset, training policy, and number of training steps, then start cloud or local training. After training completes, pull or register the model locally. A generated model must still be validated on the robotic arm; training completion alone does not mean the task is successful.
- Run the model and validate its performance. Go to Model Run / Inference, confirm that the robotic arm, camera layout, and task environment match those used during collection, select the corresponding model, and begin with a single short test. Continuously observe the robotic arm while it is running. Stop immediately if its motion deviates from the target or approaches a dangerous position.
| Stage | What to complete | Result | Next step |
|---|---|---|---|
| Device setup | Select the device model, bind the leader arm, follower arm, and cameras, then complete calibration and teleoperation verification. | A device configuration ready for control and data collection. | Collect demonstrations with the current configuration. |
| Data collection | Demonstrate the target task with the leader arm and record images, joint states, and action sequences. | A raw dataset. | Inspect collection quality on the Dataset page. |
| Dataset management | Inspect videos, frame rates, joint curves, and episodes; delete failed or abnormal data and merge datasets when necessary. | A stable, task-consistent training dataset. | Select the dataset for training. |
| Training | Select the dataset, training policy, and training steps, then start cloud or local training. | A model that can be loaded on the Run page. | Validate the model on the physical robotic arm. |
| Model run / inference | Select a model that matches the current device and task, then begin with a single short run. | A successful, partially successful, or failed validation result. | Keep the configuration, collect supplemental data, or retrain based on the result. |
- For teleoperation only or recording fixed actions: complete device setup, calibration, and teleoperation verification. Cameras are optional when visual features are not used.
- To train a model and let the robotic arm execute a task autonomously: complete data collection, dataset inspection, training, and model execution in sequence. Each stage uses the result of the previous stage, so confirm that the current stage has completed successfully before continuing.
- Continue based on execution results: after a successful run, keep the current device, data, and training configuration and gradually increase the number of tests. For partial success, collect more data for positions, poses, or process stages where the model tends to fail. For failure, first check device connections, camera positions, and the task environment, then inspect the dataset for failed or abnormal samples; correct the issues and retrain.